---

HDRT-TIR-DE: A Large-Scale Reference Dataset for TIR Image Restoration

This is part of the official repository of the paper Self-supervised Diffusion-guided Hallucination-free Thermal Infrared Image Denoising, 2026.
 

Our approach uses diffusion-based image enhancement and realistic TIR image degradation to generate image pairs for supervised learning (a) and leverages remarkable visual quality of diffusion models (c) without suffering from hallucinations (d-e).

 

Overview

The HDRT-TIR-DE dataset is part of the work described in "Self-supervised Diffusion-guided Hallucination-free Thermal Infrared Image Denoising", F. Hazebrouck, A. Schock-Schmidtke, N. Stuhrmann, J. Fottner, M. Teutsch (2026). The paper is available here. In the name HDRT-TIR-DE, TIR stands for Thermal InfraRed and DE for Diffusion-Enhanced, while HDRT is the name of the dataset the HDRT-TIR-DE build upon.

The HDRT-TIR-DE dataset is a large-scale reference thermal infrared single-image dataset designed to serve as clean reference for self-supervised training schemes.
The dataset builds upon the thermal part of the HDRT dataset, which was enhanced with an image-restoration diffusion model (StableSR) to achieve better perceptual quality than any existing real TIR dataset. Combined with the diverse scenes and high image resolution inherited from the original HDRT-TIR dataset, the HDRT-TIR-DE dataset is particularly well suited for self-supervised training of image restoration networks. The specific restoration task can be determined by the degradation model used for generating the LQ counterparts from the clean images. In our work, we placed focus on TIR sensor-noise removal.

This dataset should contribute to filling the current lack of clean TIR reference images, due to imperfections in real TIR imagers that inevitably introduce noise in every captured real TIR image.

 

Files

• HDRT-TIR-DE.zip contains the TIR images of the HDRT-TIR-DE with characteristics reported in the next section.

• train_split.txt contains the file-names of the images belonging to the train split of the HDRT-TIR-DE dataset. The split contains 80% of the dataset images, is the same as the train split of the original HDRT dataset and was used for training the models of the paper.

• val_split.txt contains the file-names of the images belonging to the validation split of the HDRT-TIR-DE dataset. The split contains 10% of the dataset images and was obtained by halving the validation split of the original HDRT dataset.

• test_split.txt contains the file-names of the images belonging to the test split of the HDRT-TIR-DE dataset. The split contains 10% of the dataset images and is the other half of the validation split of the original HDRT dataset. All results reported in the paper are computed on this split.

• noisy_test_set_cropped_center.zip contains the noisy images of the test set. They were obtained by using the noise-generator once on the HQ images of the test-split and are given here for comparability and reproducibility of the results reported in the paper.

• clean_test_set_cropped_center.zip contains the clean target images of the test set. They are a center-cropped version of the test-split and are given here for comparability and reproducibility of the results reported in the paper.

 

Dataset Characteristics

The images from the HDRT-TIR-DE dataset have a resolution of 1280×960, a bit depth of 8 and capture wavelengths from 8 to 14 μm (LWIR). This corresponds to a captured temperature range from -20°C to 100°C.
According to [1], the images were captured across three distinct seasons over a period of six months and in eight different cities located at various latitudes, ensuring a wide range of environmental conditions and thermal contexts. The images include urban and rural landscapes, natural scenery, trees, parks, vehicles, pedestrians, and buildings, under well exposed, overexposed, and underexposed conditions.
The original images from the HDRT dataset [1] were captured with an Optris PI 640i (640×480) camera. More details about the dataset can be found in the paper.  

 

Citation

If you use the HDRT-TIR-DE dataset, please cite our work:

@article{hazebrouck_self-supervised_2026,
    title = {Self-supervised Diffusion-guided Hallucination-free Thermal Infrared Image Denoising},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
    author = {Hazebrouck, Félix and Schock-Schmidtke, Alexander and Stuhrmann, Norbert and Fottner, Johannes and Teutsch, Michael},
    year = {2026},
    pages = {7091--7101},
}

 

Licence and Copyrights

This dataset is derived from the HDRT dataset.
According to the license of the HDRT Dataset, we include the original dataset authors, a copyright notice, a link to original dataset, a link to license, and a statement of modifications.

Original authors:
© Jingchao Peng, Thomas Bashford-Rogers, Francesco Banterle, Haitao Zhao, and Kurt Debattista

Original dataset:
https://huggingface.co/datasets/jingchao-peng/HDRTDataset

License:
Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International
https://spdx.org/licenses/CC-BY-NC-SA-4.0

Modifications:
We applied a diffusion-based enhancement model to improve the perceptual quality of the thermal images.

This dataset is distributed under the same license (CC-BY-NC-SA-4.0).

 

References

[1] Jingchao Peng, Thomas Bashford-Rogers, Francesco Banterle, Haitao Zhao, and Kurt Debattista. HDRT: A large-scale dataset for infrared-guided HDR imaging. Elsevier Information Fusion, 120, 2025